AI harness

Point Cursor, Claude Code, or Codex at your codebase and it simply reaches for two patterns: the most common one from its training data, and the one already most common in your repo, usually legacy code. An AI harness fixes that.

What you’re aiming for

• Agent output following the architecture your team agreed on, keeping the codebase consistent as it grows.
• Every engineer shipping at the level of your strongest one, with hard-won conventions encoded and shared across the team.
• AI spend you can account for: clarity on which tools and workflows pay off, and tokens going where they earn their keep.
• One playbook for the whole team, so engineers, designers, and writers pull from the same source of truth.

The AI harnesses Evil Martians build

An AI harness encodes your conventions, architecture, and guardrails as deterministic tools, skills, evals, and telemetry. We build it for your stack, whether you ship Rails, TypeScript, Go, or Python, so agents produce code your senior engineers would approve.

• Skills and a shared marketplace: your team’s best practices and unique expertise, encoded once and reused by everyone. Designers encode brand guidelines, writers distribute tone of voice, and business folks share ICP and research skills with the whole team.
• Guardrails that hold as you grow: when you ship a new design system or architectural pattern, agents default to the legacy version because there’s more of it in the code. We ship deterministic linters and hooks that catch the divergence and tell the agent how to follow the new convention.
• Evals: tests that confirm skills behave correctly and fire on the right requests.
• Observability: telemetry and a dashboard into skill usage and gaps, built on AgentPrism, our open source agent-trace visualization library.
• Cost control: telemetry on where your AI tokens go, with expensive AI calls swapped for cheaper deterministic checks where the task allows.

What we’ve already shipped this way

Evil Martians cut the Rails test suite and CI time in half for the Rails unicorn Whop, then scoped a full AI harness for their codebase: automated linting on skills, evals that confirmed skills behaved correctly and fired on the right requests, telemetry into skill usage and gaps, and gating that caught regressions before they shipped.

Evil Martians added LLM safeguards to Playbook, cutting engineering managers’ PR review time by filtering routine issues before a human reads the pull request.

For the fastest-growing AI support startup, Evil Martians run forward-deployed engineering at scale.

Internally, Evil Martians run the same harness architecture across our 50-person distributed team. The skills, hooks, conventions, and dashboards we use every day are the same ones we bring into client codebases.

What it takes to start

A 2-week sprint is $16,000. A senior Martian engineer works alongside your team for two weeks. We read your Git history, strategize with your principal engineers, and ship the first working version of the harness.

From you: access to your codebase and Git history, plus time with your principal engineers to set direction.

From us: a senior Martian for two weeks at $16,000. Extendable at $8,000 per week.